Chapter 9

Three-Dimensional Lookup Table

with Interpolation

Color space transformation using a 3D lookup table (LUT) with interpolation is

used to correlate the source and destination color values in the lattice points of
a 3D table, where nonlattice points are interpolated by using the nearest lattice
points. This method has been used in many applications with quite satisfactory
results, and incorporated into the ICC profile standard.1
In this chapter, the structure of the 3D-LUT approach is discussed and the math-
ematical formulations of interpolation methods are given. These methods are tested
using several sets of data points. The similarities and differences of these interpo-
lation methods are discussed.

9.1 Structure of 3D Lookup Table

The 3D lookup method consists of three parts—packing (or partition), extraction

(or find), and interpolation (or computation).2 Packing is a process that partitions
the source space and selects sample points for the purpose of building a lookup
table. The extraction step is aimed at finding the location of the input pixel and
extracting color values of the nearest lattice points. The last step is interpolation
where the input signals and the extracted lattice points are used to calculate the
destination color specifications for the input point.

9.1.1 Packing

Packing is a process that divides the domain of the source space and populates it
with sample points to build the lookup table. Generally, the table is built by an
equal step sampling along each axis of the source space, as shown in Fig. 9.1,
of a five-level LUT. This will give (n − 1)3 cubes and n3 lattice points, where
n is the number of levels. The advantage of this arrangement is that it implicitly
supplies information about which cell is next to which. Thus, one needs only to
store the starting point and the spacing for each axis. Generally, a matrix of n3
color patches at the lattice points of the source space is made and the destination
color specifications of these patches are measured. The corresponding values from
the source and destination spaces are tabulated into a lookup table.
151
152 Computational Color Technology

Figure 9.1 A uniformly spaced five-level 3D packing.

Nonuniformly spaced LUTs are also used extensively. They lose the simplicity
of the implementation edges, but gain the versatility and conversion accuracy as
discussed in Section 9.7.

9.1.2 Extraction

Nonlattice points are interpolated by using the nearest lattice points. This is the
step where the extraction performs a search to select the lattice points necessary
for computing the destination specification of the input point. A well-packed space
can make the search simpler. In an 8-bit integer environment, for example, if the
axis is divided into 2j equal sections where j is an integer smaller than 8, then the
nearest lattice points are given in the most significant j bits (MSBj ) of the input
color signals. In other words, the input point is bounded between the lattice points
p(MSBj ) and p(MSBj + 1). To find the point of interest requires the computer
operations of the masking byte and shifting bits that are significantly faster than
the comparison operations. For unequally spaced packing, a series of comparisons
are needed to locate the nearest lattice points.
Further search within the cube may be needed, depending on the interpolation
technique employed to compute the color values of nonlattice points. There are two
interpolation methods: the geometrical method and cellular regression. Cellular re-
gression does not require a search within the cube. All geometric interpolations
Three-Dimensional Lookup Table with Interpolation 153

except the trilinear approach require a search mechanism to find the subdivided
structure where the point resides. These search mechanisms are inequality com-
parisons.

9.1.3 Interpolation

Interpolation uses the input signals and the extracted lattice points that contain the
destination specifications to calculate the destination color specifications for the
input point. Interpolation techniques are mathematical computations that employ
geometrical relationships or cellular regression. Geometrical interpolations exploit
the ways of subdividing a cube. There are four geometrical interpolations, trilin-
ear, prism, pyramid, and tetrahedral. The first 3D interpolation that appeared in
the literature is the trilinear interpolation, disclosed in a 1974 British patent by
Pugsley.3 The prism scheme was published in 1992 by Kanamori, Fumoto, and
Kotera.4 This prism architecture has been made into a single-chip color proces-
sor by Matsushita Research Institute Tokyo, Inc.5,6 The pyramid interpolation was
patented by Flanklin in 1982.7 The idea of linear interpolation using tetrahedral
segmentation of a cube was published as early as 1967 by Iri.8 A similar concept
of a linear interpolation by searching for the nearest four neighbors that enclose
the point of interest and form a tetrahedron was applied to compute dot areas of
color scanners by Korman and Yule in 1971.9 The application of tetrahedral inter-
polation to color-space transformation was later patented by Sakamoto and Itooka
in U.S. Patent No. 4,275,413 (1981) and related worldwide patents.10–12 The ex-
tensive activities in developing and patenting interpolation techniques during the
1980s show the importance of the technique, the desire of dominating market share
by the manufacturers, and the subsequent financial stakes.

9.2 Geometric Interpolations

Basically, 3D interpolation is the multiple application of the linear interpolation;

therefore, we start with the linear interpolation, then extend to 2D (bilinear) and
3D (trilinear) interpolations. A linear interpolation is depicted in Fig. 9.2; a point
p on the curve between the lattice points p0 and p1 is to be interpolated. The
interpolated value pc (x) is linearly proportional to the ratio of (x − x0 )/(x1 − x0 ),
where (x1 − x0 ) is the projected length of the line segment connecting points p0
and p1 , and (x − x0 ) is the projected distance of the line connecting points p and
p0 .

pc (x) = p(x0 ) + [(x − x0 )/(x1 − x0 )][p(x1 ) − p(x0 )]. (9.1)

As shown in Eq. (9.1), the major computational operation in the interpolation is to

calculate the projected distances. In view of the implementation, using a uniform
8-bit LUT, the projected distance at each axis of an input is simply p(LSB8−j ),
154 Computational Color Technology

Figure 9.2 Linear interpolation.

where the LSBs are the least significant bits. This is a simple byte-masking op-
eration that significantly lowers the computational cost and increases speed. The
interpolation error is given as

δ = p(x) − pc (x). (9.2)

9.2.1 Bilinear interpolation

In two dimensions, we have a function of two variables p(x, y) and four lattice
points p00 (x0 , y0 ), p01 (x0 , y1 ), p10 (x1 , y0 ), and p11 (x1 , y1 ) as shown in Fig. 9.3.
To obtain the value for point p, we first hold y0 constant and apply the linear
interpolation on lattice points p00 and p10 to obtain p0 .

p0 = p00 + (p10 − p00 )[(x − x0 )/(x1 − x0 )]. (9.3)

Similarly, we calculate p1 by keeping y1 constant.

p1 = p01 + (p11 − p01 )[(x − x0 )/(x1 − x0 )]. (9.4)

After obtaining p0 and p1 , we again apply the linear interpolation to them by keep-
ing x constant.

p(x, y) = p0 + (p1 − p0 )[(y − y0 )/(y1 − y0 )]. (9.5)

Three-Dimensional Lookup Table with Interpolation 155

Figure 9.3 Bilinear interpolation.

Substituting Eqs. (9.3) and (9.4) into Eq. (9.5), we obtain

p(x, y) = p00 + (p10 − p00 )[(x − x0 )/(x1 − x0 )]

+ (p01 − p00 )[(y − y0 )/(y1 − y0 )]
+ (p11 − p01 − p10 + p00 )[(x − x0 )/(x1 − x0 )]
× [(y − y0 )/(y1 − y0 )]. (9.6)

9.2.2 Trilinear interpolation

The trilinear equation is derived by applying the linear interpolation seven times
(see Fig. 9.4); three times each to determine the points p1 and p0 as illustrated
in the 2D bilinear interpolation, then one more time to compute the point p. The
general expression for the trilinear interpolation is given in Eq. (9.7).

p(x, y, z) = c0 + c1 x + c2 y + c3 z + c4 xy
+ c5 yz + c6 zx + c7 xyz, (9.7a)

where x, y, and z are the relative distances of the point with respect to the
starting point p000 in the x, y, and z directions, respectively, as shown in Eq. (9.7b).

x = (x − x0 )/(x1 − x0 ); y = (y − y0 )/(y1 − y0 ); z = (z − z0 )/(z1 − z0 ).

(9.7b)
156 Computational Color Technology

Figure 9.4 Trilinear interpolation.

Coefficients cj are determined from the values of the vertices.

c0 = p000 ; c1 = (p100 − p000 ); c2 = (p010 − p000 );

c3 = (p001 − p000 ); c4 = (p110 − p010 − p100 + p000 );
c5 = (p011 − p001 − p010 + p000 ); c6 = (p101 − p001 − p100 + p000 );
c7 = (p111 − p011 − p101 − p110 + p100 + p001 + p010 − p000 ). (9.7c)

Equation (9.7a) can be written in vector-matrix form as

p = C T Q1 or p = QT1 C, (9.7d)

where C is the vector of coefficients,

C = [c0 c1 c2 c3 c4 c5 c6 c7 ]T , (9.8)

and Q1 is the vector of distances related to x, y, and z.

Q1 = [1 x y z xy yz zx xyz]T . (9.9)

Three-Dimensional Lookup Table with Interpolation 157

Note that the sizes of vectors Q1 and C must be the same. The coefficients C can
be put into vector-matrix form as shown in Eq. (9.10a) by expanding Eq. (9.7c).
    
c0 1 0 0 0 0 0 0 0 p000
 c1   −1 0 0 0 1 0 0 0   p001 
    
 c2   −1 0 1 0 0 0 0 0   p010 
    
 3 
c −1 1 0 0 0 0 0 0   p011 
 =  , (9.10a)
 c4   1 0 −1 0 −1 0 1 0   p100 
    
 c5   1 −1 −1 1 0 0 0 0   p101 
 c   1 −1 0 0 −1 1 0 0   p110 
6
c7 −1 1 1 −1 1 −1 −1 1 p111

or

C = B 1P , (9.10b)

where vector P is a collection of vertices,

P = [p000 p001 p010 p011 p100 p101 p110 p111 ]T , (9.11)

and the matrix B 1 , given in Eq. (9.10a), represents a matrix of binary constants,
having a size of 8 × 8.
Substituting Eq. (9.10b) into Eq. (9.7d), we obtain the vector-matrix expression
for calculating the destination color value of point p.

p = C T Q1 = P T B T1 Q1 , (9.12a)
p = QT1 C = QT1 B 1 P . (9.12b)

Equation (9.12) is exactly the same as Eq. (9.7), only expressed differently. There
is no need for a search mechanism to find the location of the point because the
cube is used as a whole. But the computation cost, using all eight vertices and
having eight terms in the equation, is higher than other 3D geometric interpola-
tions.

9.2.3 Prism interpolation

If one cuts a cube diagonally into two halves as shown in Fig. 9.5, one gets two
prism shapes. A search mechanism is needed to locate the point of interest. Because
there are two symmetric structures in the cube, a simple inequality comparison is
sufficient to determine the location: if x > y, then the point is in Prism 1,
otherwise the point is in Prism 2. For x = y, the point is laid on the diagonal
plane, and either prism can be used for interpolation.
158 Computational Color Technology

Figure 9.5 Prism interpolation.

The equation has six terms and uses six vertices of the given prism for compu-
tation. Equation (9.13) gives prism interpolation in vector-matrix form.
 
p1
p2

p000 (p100 − p000 ) (p110 − p100 ) (p001 − p000 )
=
p000 (p110 − p010 ) (p010 − p000 ) (p001 − p000 )

(p101 − p001 − p100 + p000 ) (p111 − p101 − p110 + p100 )
(p111 − p011 − p110 + p010 ) (p011 − p001 − p010 + p000 )
 
1
 x 
 
 y 
× . (9.13)
 z 
 xz 
yz

By setting

Q2 = [1 x y z xz yz]T , (9.14)

Eq. (9.13) can be expressed in vector-matrix form as given in Eq. (9.15).

p1 = P T B T21 Q2 = QT2 B 21 P , (9.15a)

p2 = P T B T22 Q2 = QT2 B 22 P , (9.15b)
Three-Dimensional Lookup Table with Interpolation 159

where vector P is given in Eq. (9.11), and B 21 and B 22 are binary matrices, having
a size of 6 × 8, given as follows:

 
1 0 0 0 0 0 0 0
 −1 0 0 1 0 0 0 0
 
 0 0 0 0 −1 0 1 0
B 21 =  ,
 −1 1 0 0 0 0 0 0
 1 −1 0 0 −1 1 0 0
0 0 0 0 1 −1 −1 1
 
1 0 0 0 0 0 0 0
 0 0 −1 0 0 0 1 0
 
 −1 0 1 0 0 0 0 0
B 22 =  .
 −1 1 0 0 0 0 0 0
 0 0 1 −1 0 0 −1 1
1 −1 −1 1 0 0 0 0

The location of the data point is determined by the following IF-ELSE construct:

If x > y, p = p1 ,
else p = p2 .

9.2.4 Pyramid interpolation

For pyramid interpolation, the cube is sliced into three pieces; each one takes a face
as the pyramid base, having its corners connected to a vertex in the opposite side
as the apex (see Fig. 9.6). A search is required to locate the point of interest. The
equation has five terms and uses five vertices of the given pyramid to compute the
value.
Equation (9.16) gives the vector-matrix form of the pyramid interpolation.

Figure 9.6 Pyramid interpolation.